Rowing0914 / dm_control2gym

Introduction

I simplified martinseilair's dm_control2gym as well as made it compatible with the latest MuJoCo library(e.g., 2.0.0). .... yeah,, I will upload the documents when I have time.

Installation

$ pip install dm_control2gym

Dependencies

• Please check the requirements.txt or just pip install -r requirements.txt

Usage

• MDP tasks: env returns a state of a robot at each time-step

  import itertools
  from dm_control2gym.util import make_dm2gym_env_state
  
  env = make_dm2gym_env_state(env_name="cheetah_run")
  
  state = env.reset()
  print("State shape: ", state.shape)
  
  total_reward = 0
  
  for t in itertools.count():
      action = env.action_space.sample()
      state, reward, done, _ = env.step(action)
      total_reward += reward
  
      if done: break
  
  env.close()
  print("Total Reward: {}".format(total_reward))

• POMDP tasks: env returns a raw image observation at each time-step

  import itertools
  from dm_control2gym.util import make_dm2gym_env_obs
  from dm_control2gym.recorder import Monitor
  
  env = make_dm2gym_env_obs(env_name="cheetah_run", num_repeat_action=1)
  env = Monitor(env=env, directory="./log", force=True)
  
  obs = env.reset()
  print("Obs shape: ", obs.shape)
  
  total_reward = 0
  
  env.record_start()
  env.reset()
  for t in itertools.count():
      action = env.action_space.sample()
      obs, reward, done, _ = env.step(action)
      total_reward += reward
  
      if done: break
  
  env.record_end()
  env.close()
  print("Total Reward: {}".format(total_reward))

Videos

For more details

Please refer to martinseilair's dm_control2gym